root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-26 03:14:35 +00:00
Code Issues Actions 7 Packages Projects Releases Wiki Activity

[SYCL] fix error when set main gpu to non-zero (#5901)

Browse Source 
* fix error when set main gpu to non-zero

* fix delete condition
This commit is contained in:
Neo Zhang Jianyu 2024-03-07 16:34:31 +08:00 committed by GitHub
parent e04e04f8fa
commit ceca1aef07
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 119 additions and 76 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

102
ggml-sycl.cpp
View File

@ -3559,12 +3559,31 @@ class sycl_gpu_mgr {
int work_group_size = 0; int work_group_size = 0;
std::string gpus_list = ""; std::string gpus_list = "";
/*
Use all GPU with same top max compute units
*/
sycl_gpu_mgr() { sycl_gpu_mgr() {
detect_sycl_gpu_list_with_max_cu(); detect_sycl_gpu_list_with_max_cu();
get_allow_gpus(); get_allow_gpus();
create_context_with_gpus(); create_context_with_gpus();
} }
/*
Use the assigned GPU as only one
*/
sycl_gpu_mgr(int main_gpu_id) {
sycl::device device = dpct::dev_mgr::instance().get_device(main_gpu_id);
dpct::device_info prop;
dpct::get_device_info(prop, device);
gpus.push_back(main_gpu_id);
devices.push_back(device);
work_group_size = prop.get_max_work_group_size();
max_compute_units = prop.get_max_compute_units();
get_allow_gpus();
create_context_with_gpus();
}
void create_context_with_gpus() { void create_context_with_gpus() {
sycl::context ctx = sycl::context(devices); sycl::context ctx = sycl::context(devices);
assert(gpus.size() > 0); assert(gpus.size() > 0);
@ -3580,7 +3599,7 @@ class sycl_gpu_mgr {
gpus_list += std::to_string(gpus[i]); gpus_list += std::to_string(gpus[i]);
gpus_list += ","; gpus_list += ",";
} }
if (gpus_list.length() > 2) { if (gpus_list.length() > 1) {
gpus_list.pop_back(); gpus_list.pop_back();
} }
} }
@ -3629,8 +3648,8 @@ class sycl_gpu_mgr {
if (gpus[i] == id) if (gpus[i] == id)
return i; return i;
} }
assert(false); printf("miss to get device index by id=%d\n", id);
return -1; GGML_ASSERT(false);
} }
int get_next_index(int id) { int get_next_index(int id) {
@ -3639,8 +3658,7 @@ class sycl_gpu_mgr {
if (gpus[i] == id) if (gpus[i] == id)
return i; return i;
} }
assert(false); GGML_ASSERT(false);
return -1;
} }
}; };
@ -3649,6 +3667,7 @@ static int g_device_count = -1;
static int g_all_sycl_device_count = -1; static int g_all_sycl_device_count = -1;
static int g_main_device = -1; static int g_main_device = -1;
static int g_main_device_id = -1; static int g_main_device_id = -1;
static bool g_ggml_backend_sycl_buffer_type_initialized = false;
static std::array<float, GGML_SYCL_MAX_DEVICES> g_default_tensor_split = {}; static std::array<float, GGML_SYCL_MAX_DEVICES> g_default_tensor_split = {};
@ -13225,7 +13244,7 @@ void ggml_backend_sycl_print_sycl_devices() {
} }
void print_gpu_device_list() { void print_gpu_device_list() {
fprintf(stderr, "detect %d SYCL GPUs: [%s] with Max compute units:%d\n", fprintf(stderr, "detect %d SYCL GPUs: [%s] with top Max compute units:%d\n",
g_sycl_gpu_mgr->get_gpu_count(), g_sycl_gpu_mgr->get_gpu_count(),
g_sycl_gpu_mgr->gpus_list.c_str(), g_sycl_gpu_mgr->gpus_list.c_str(),
g_sycl_gpu_mgr->max_compute_units); g_sycl_gpu_mgr->max_compute_units);
@ -13264,23 +13283,6 @@ void ggml_init_sycl() try {
#else #else
fprintf(stderr, "%s: GGML_SYCL_F16: no\n", __func__); fprintf(stderr, "%s: GGML_SYCL_F16: no\n", __func__);
#endif #endif
if (CHECK_TRY_ERROR(g_all_sycl_device_count =
dpct::dev_mgr::instance().device_count()) != 0) {
initialized = true;
g_sycl_loaded = false;
return;
}
GGML_ASSERT(g_all_sycl_device_count <= GGML_SYCL_MAX_DEVICES);
ggml_backend_sycl_print_sycl_devices();
if (!g_sycl_gpu_mgr) g_sycl_gpu_mgr = new sycl_gpu_mgr();
g_device_count = g_sycl_gpu_mgr->get_gpu_count();
g_work_group_size = g_sycl_gpu_mgr->work_group_size;
print_gpu_device_list();
int64_t total_vram = 0;
/* NOT REMOVE, keep it for next optimize for XMX. /* NOT REMOVE, keep it for next optimize for XMX.
#if defined(SYCL_USE_XMX) #if defined(SYCL_USE_XMX)
@ -13289,6 +13291,29 @@ void ggml_init_sycl() try {
fprintf(stderr, "%s: SYCL_USE_XMX: no\n", __func__); fprintf(stderr, "%s: SYCL_USE_XMX: no\n", __func__);
#endif #endif
*/ */
if (CHECK_TRY_ERROR(g_all_sycl_device_count =
dpct::dev_mgr::instance().device_count()) != 0) {
initialized = true;
g_sycl_loaded = false;
return;
}
GGML_ASSERT(g_all_sycl_device_count <= GGML_SYCL_MAX_DEVICES);
ggml_backend_sycl_print_sycl_devices();
if (!g_sycl_gpu_mgr) g_sycl_gpu_mgr = new sycl_gpu_mgr();
print_gpu_device_list();
initialized = true;
g_sycl_loaded = true;
}
g_device_count = g_sycl_gpu_mgr->get_gpu_count();
g_work_group_size = g_sycl_gpu_mgr->work_group_size;
int64_t total_vram = 0;
for (int id = 0; id < GGML_SYCL_MAX_DEVICES; ++id) { for (int id = 0; id < GGML_SYCL_MAX_DEVICES; ++id) {
g_device_caps[id].vmm = 0; g_device_caps[id].vmm = 0;
g_device_caps[id].device_id = -1; g_device_caps[id].device_id = -1;
@ -13331,10 +13356,6 @@ void ggml_init_sycl() try {
// create sycl handle // create sycl handle
SYCL_CHECK(CHECK_TRY_ERROR(g_sycl_handles[i] = stream)); SYCL_CHECK(CHECK_TRY_ERROR(g_sycl_handles[i] = stream));
} }
initialized = true;
g_sycl_loaded = true;
}
} }
catch (sycl::exception const &exc) { catch (sycl::exception const &exc) {
std::cerr << exc.what() << "Exception caught at file:" << __FILE__ std::cerr << exc.what() << "Exception caught at file:" << __FILE__
@ -16732,22 +16753,24 @@ static ggml_backend_buffer_type_i ggml_backend_sycl_buffer_type_interface = {
/* .is_host = */ nullptr, /* .is_host = */ nullptr,
}; };
ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device) { ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device_index) {
if (device_index>=g_device_count or device_index<0) {
printf("ggml_backend_sycl_buffer_type error: device_index:%d is out of range [0, %d], miss to call ggml_backend_sycl_set_single_device()\n",
device_index, g_device_count-1);
GGML_ASSERT(device_index<g_device_count);
}
static struct ggml_backend_buffer_type ggml_backend_sycl_buffer_types[GGML_SYCL_MAX_DEVICES]; static struct ggml_backend_buffer_type ggml_backend_sycl_buffer_types[GGML_SYCL_MAX_DEVICES];
static bool ggml_backend_sycl_buffer_type_initialized = false; if (!g_ggml_backend_sycl_buffer_type_initialized) {
if (!ggml_backend_sycl_buffer_type_initialized) {
for (int i = 0; i < g_device_count; i++) { for (int i = 0; i < g_device_count; i++) {
ggml_backend_sycl_buffer_types[i] = { ggml_backend_sycl_buffer_types[i] = {
/* .iface = */ ggml_backend_sycl_buffer_type_interface, /* .iface = */ ggml_backend_sycl_buffer_type_interface,
/* .context = */ new ggml_backend_sycl_buffer_type_context{i, GGML_SYCL_NAME + std::to_string(g_sycl_gpu_mgr->gpus[i])}, /* .context = */ new ggml_backend_sycl_buffer_type_context{i, GGML_SYCL_NAME + std::to_string(g_sycl_gpu_mgr->gpus[i])},
}; };
} }
ggml_backend_sycl_buffer_type_initialized = true; g_ggml_backend_sycl_buffer_type_initialized = true;
} }
return &ggml_backend_sycl_buffer_types[device_index];
return &ggml_backend_sycl_buffer_types[device];
} }
// sycl split buffer type // sycl split buffer type
@ -17496,6 +17519,17 @@ GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id) {
return g_sycl_gpu_mgr->get_index(device_id); return g_sycl_gpu_mgr->get_index(device_id);
} }
GGML_API GGML_CALL void ggml_backend_sycl_set_single_device(int main_gpu_id) {
GGML_ASSERT(main_gpu_id<g_all_sycl_device_count);
printf("ggml_backend_sycl_set_single_device: use single device: %d\n", main_gpu_id);
if (g_sycl_gpu_mgr) {
delete g_sycl_gpu_mgr;
}
g_sycl_gpu_mgr = new sycl_gpu_mgr(main_gpu_id);
ggml_init_sycl();
g_ggml_backend_sycl_buffer_type_initialized = false;
}
extern "C" int ggml_backend_sycl_reg_devices(); extern "C" int ggml_backend_sycl_reg_devices();
int ggml_backend_sycl_reg_devices() { int ggml_backend_sycl_reg_devices() {

1
ggml-sycl.h
View File

@ -28,6 +28,7 @@ GGML_API GGML_CALL int ggml_backend_sycl_get_device_count();
GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split); GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
GGML_API GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total); GGML_API GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id); GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id);
GGML_API GGML_CALL void ggml_backend_sycl_set_single_device(int main_gpu);
#ifdef __cplusplus #ifdef __cplusplus
} }

16
llama.cpp
View File

@ -3750,6 +3750,14 @@ static bool llm_load_tensors(
model.main_gpu = main_gpu; model.main_gpu = main_gpu;
model.n_gpu_layers = n_gpu_layers; model.n_gpu_layers = n_gpu_layers;
#ifdef GGML_USE_SYCL
if (split_mode == LLAMA_SPLIT_MODE_NONE) {
ggml_backend_sycl_set_single_device(main_gpu);
//SYCL use device index (0, 1, 2), instead if device id.
main_gpu = ggml_backend_sycl_get_device_index(main_gpu);
}
#endif
const int64_t n_layer = hparams.n_layer; const int64_t n_layer = hparams.n_layer;
const int64_t i_gpu_start = std::max((int64_t) hparams.n_layer - n_gpu_layers, (int64_t) 0); const int64_t i_gpu_start = std::max((int64_t) hparams.n_layer - n_gpu_layers, (int64_t) 0);
@ -12260,13 +12268,13 @@ struct llama_context * llama_new_context_with_model(
ctx->backends.push_back(backend); ctx->backends.push_back(backend);
} else { } else {
// LLAMA_SPLIT_LAYER requires a backend for each GPU // LLAMA_SPLIT_LAYER requires a backend for each GPU
int id_list[GGML_SYCL_MAX_DEVICES];
ggml_sycl_get_gpu_list(id_list, GGML_SYCL_MAX_DEVICES);
for (int i = 0; i < ggml_backend_sycl_get_device_count(); ++i) { for (int i = 0; i < ggml_backend_sycl_get_device_count(); ++i) {
int device_id = id_list[i];
ggml_backend_t backend = ggml_backend_sycl_init(i); ggml_backend_t backend = ggml_backend_sycl_init(i);
if (backend == nullptr) { if (backend == nullptr) {
LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d (index %d)backend\n", __func__, device_id, i); int id_list[GGML_SYCL_MAX_DEVICES];
ggml_sycl_get_gpu_list(id_list, GGML_SYCL_MAX_DEVICES);
LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d (index %d)backend\n", __func__, id_list[i], i);
llama_free(ctx); llama_free(ctx);
return nullptr; return nullptr;
} }